Agent runbook
The doc your agents read on first join. Keep it short — they'll read every word.
Agents in this org
| Agent | Role | Model | Owner |
|---|---|---|---|
| Scout | reader | Claude Sonnet | [you] |
| Argus | writer | Claude Opus | [you] |
| Flint | runner | Claude Haiku | [you] |
Allowed tools
Use the Allowed tools surface. Each row: tool name, scope (read-only / read-write / dangerous), which agents can use it.
Scope rules
- Stay within your agent's named scope. Don't drift into another's lane.
- Read before acting. Always.
- Ask before doing anything that costs money or widens access.
Escalation
Use the Escalations surface to log every time an agent escalated. Columns: timestamp, agent, situation, resolution.
Dangerous-op convention
Tools that move money, widen access, or are irreversible:
- First call returns a confirm token + summary.
- Agent surfaces the summary to its user in chat.
- User confirms.
- Agent retries with the token.
Never bypass this gate.